02 / Case studies
Selected work in autonomous AI, product ownership, workflow transformation, analytics, and operations.
01 / An ongoing investment experiment
Human vs. Machine. Same market. No excuses.
Performance observations are not yet available.
Loading the dated experiment snapshot…
Open the full experiment02 / Product ownership
Cycling instructors should not have to replay every song and mark every transition before they can design a class.
GRADE.Build the ride.I conceived GRADE as a browser prototype for turning song structure into an editable cycling-class plan. That original product vision became the foundation for a five-person Agentic AI course project.
As Founder and Product Owner, I set the product direction, translate cycling-instructor needs into requirements, prioritize features, test workflows, and guide team decisions. GRADE 2.0 is collaborative work; I do not claim sole implementation authorship.
The current full-stack application accepts audio files or YouTube links, analyzes song structure on a Python server, lets instructors edit and order effort blocks, saves programs, builds multi-song class plans, and runs the class with cues and countdowns.
Natural-language class generation is an agreed next capability: instructors will describe the ride they want, then refine the generated plan rather than starting from a blank timeline.
Eight build milestones and extensive automated tests are documented in the public repository. Real-device testing, public hosting, mobile background playback, storage cleanup, and production hardening remain open.
03 / AI workflow transformation
I was not handed a requirements document. I sat with underwriters, watched the work, tried it myself, and found where judgment was being displaced by repetitive research and document handling.
Condo referrals required underwriters to search multiple broker documents for specific fields, manually transfer information into a referral format, and conduct broad public-record research. The work was time-consuming and varied by user.
I owned problem discovery, workflow design, prompt architecture, historical-case testing, feedback collection, iteration, training, documentation, and handoff. The AI skills were configuration-driven rather than traditionally coded.
Four Chubb AI workflows: condo intelligence research, referral document extraction, referral quality assurance, and a coach for scoping other AI use cases. I also created a searchable condo-rate reference, a mailroom inventory tracker, and an office seating audit workbook.
I ran the tools against historical cases and compared outputs with completed human referrals and research. I tracked missed risks, inconsistent formatting, false positives, and known document-reading failures.
Across test cases, the research workflows found signals including fires, shootings, lawsuits, resident complaints, and foreign-language local reporting that had not always appeared in the original research. Every finding required source review by an underwriter.
Some underwriters wanted a comprehensive report; others wanted to guide the investigation conversationally. I created a second, menu-driven research coach that recommended relevant paths while leaving control with the user. Adoption depended on fitting the workflow—not merely producing more information.
Status: prototypes were tested, demonstrated, documented, and handed off. Organization-wide adoption and formal time savings were not measured.
04 / Analytics
Some learners were completing a roughly 20-hour licensing course unusually quickly, then failing an exam whose eligible retakes carried a cost to the company.
I used SQL to retrieve course and practice-exam records, then analyzed completion speed, pass rates, first-attempt performance, and regional patterns in Excel.
I built formulas, PivotTables, conditional formatting, and management dashboards that allowed leaders to see where suspicious patterns concentrated.
Unusually rapid completion and pass patterns were concentrated under a small number of regional leaders in Florida. The analysis identified where follow-up was warranted; it did not establish individual intent.
I created a step-by-step guide with screenshots and an edited training video so employees in the life-insurance function could reproduce the dashboard workflow.
Original company data and dashboards are not published. A synthetic reconstruction may be added later and would be clearly labeled.
05 / Leadership & operations
A one-person campus bicycle program could only reach so many people. I built a community that could do more together.
I recruited riders through direct conversations and grassroots QR-code outreach. A three-officer team now supports rides, repair, and communication.
UGA annually identifies abandoned bicycles. We worked with the Office of Sustainability to reclaim bikes that were otherwise sitting unused, repair them, and offer affordable rentals—especially to international exchange students.
Our repair officer tracks parts and estimated avoided repair costs in a spreadsheet. Selected bike sales have helped fund the organization, and local-business collaborations have expanded our community reach.
I am leading a transition to new officers before graduation, including appointing new leadership and creating a structured selection process for the next social-media officer.
06 / Team consulting project
Ahead of the World Cup, our team assessed how five Atlanta transit stations could introduce recycling infrastructure.
We identified high-traffic receptacles, sorted and weighed their contents, calculated diversion rates, studied bin placement and foot traffic, researched other transit systems, and reviewed local recycling infrastructure.
As the team’s data analyst, I managed the audit data in Excel, built formulas and dashboards, calculated station-level diversion rates, and translated the findings into presentation-ready Canva visualizations.
We presented MARTA with recommended recycling locations, implementation considerations, and accessible poster and floor-sticker concepts for a diverse rider population.
The work demonstrated a quantified opportunity and produced a complete recommendation package. MARTA did not provide enough follow-up for us to claim that our recommendations were implemented.
View the original sustainability portfolio and presentation ↗